Multi-Scale Inshore Ship Detection Based on Feature Re-Focusing Network

Di Liu; Yan Zhang; Yan Zhao; Zhiguang Shi; Jinghua Zhang; Yu Zhang

doi:10.3788/AOS202141.2215001

Acta Optica Sinica, Volume. 41, Issue 22, 2215001(2021)

Multi-Scale Inshore Ship Detection Based on Feature Re-Focusing Network

Di Liu¹, Yan Zhang^1、*, Yan Zhao², Zhiguang Shi¹, Jinghua Zhang¹, and Yu Zhang¹

¹National Key Laboratory of Science and Technology on Automatic Target Recognition, College of Electronic Science and Technology, National University of Defense Technology, Changsha, Hunan 410073, China

²State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System, College of Electronic Science and Technology, National University of Defense Technology, Changsha, Hunan 410073, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(19)

Fig. 1. Inshore ship target in the surveillance video. (a) Size difference is large; (b) ship is cut off; (c) ships block each other; (d) ship is confused with the surrounding background

Download full size

Fig. 2. Structure of the FRN

Download full size

Fig. 3. Feature fusion strategy of the FPN

Download full size

Fig. 4. Feature fusion strategy of the MFAM

Download full size

Fig. 5. Structure of the AFRM

Download full size

Fig. 6. Structure of the CAM

Download full size

Fig. 7. Structure of the SAM

Download full size

Fig. 8. Ship target in the Seaships7000 data set. (a) Ore carrier; (b) bulk cargo carrier; (c) general cargo carrier; (d) container ship; (e) fishing ship; (f) passenger ship

Download full size

Fig. 9. Statistics of the Seaships7000 data set. (a) Size distribution; (b) percentage of quantity; (c) division of the data set

Download full size

Fig. 10. Visualization of feature maps of FRN and benchmark algorithms. (a) Overlapping ships; (b) multiple ships

Download full size

Fig. 11. PR curves of the algorithm with different IoU. (a) Value range of IoU is 0.5~0.95; (b) IoU is 0.5; (c) IoU is 0.75

Download full size

Fig. 12. PR curves of different algorithms for different types of ship. (a) Passenger ship; (b) general cargo carrier; (c) fishing ship; (d) bulk cargo carrier; (e) ore carrier; (f) container ship

Download full size

Fig. 13. Detection results of different algorithms on inshore ship targets. (a) Multi-target overlapping scene; (b) target and background confuse the scene; (c) small size target scene; (d) underlit scene

Download full size

Table 1. Structure and parameter of basic feature extraction network

View table

Table 1. Structure and parameter of basic feature extraction network

Name	Layer	Layer group	Filters number	Kernel size	Stride size	Padding size
VGG-16	Conv1_1	1	64	3	1	1
	Conv1_2	1	64	3	1	1
	Conv2_1	2	128	3	1	1
	Conv2_2	2	128	3	1	1
	Conv3_1	3	256	3	1	1
	Conv3_2
	Conv3_3
	Conv4_1	4	512	3	1	1
	Conv4_2
	Conv4_3
	Conv5_1	5	512	3	1	1
	Conv5_2
	Conv5_3
	Conv6	6	1024	3	1	1
	Conv7	7	1024	1	1	0
Extra	Extra1	8	256	1	1	0
Extra	Extra2	8	512	3	2	1

Table 2. Hyperparameters of experimental settings

View table

Table 2. Hyperparameters of experimental settings

Hyperparameter	Value
Anchor box sizes	32,64,128,256
Anchor box scales	1,1.414,0.707
Epoch	100
Optimizer	SGD
Learning rate	0.001
Learning rate decay steps	50,75
Learning rate decay rate	0.1
Weight decay	5×10^-4
Mini-batch	6

Table 3. Effect of FRS internal modules on FRN detection performance

View table

Table 3. Effect of FRS internal modules on FRN detection performance

Algorithm	AP	AP^0.5	AP^0.75	AP^s	AP^m	AP^l	FPS
FRN	0.692	0.948	0.830	0.122	0.525	0.703	25
FRN (w/o MFAM)	0.684	0.945	0.823	0.096	0.550	0.695	31
FRN (w/o AFRM)	0.686	0.952	0.826	0.092	0.536	0.699	37
FRN (w/o CAM)	0.681	0.942	0.818	0.110	0.525	0.691	29
FRN (w/o SAM)	0.682	0.946	0.824	0.153	0.508	0.693	27
FRN (w/o SAM&CAM)	0.678	0.944	0.824	0.097	0.517	0.688	34
FRN(w/o ARM)	0.683	0.945	0.824	0.101	0.487	0.695	27
RefineDet	0.674	0.946	0.815	0.134	0.505	0.685	47

Table 4. Effects of FRS internal modules on the detection performance of different types of ships

View table

Table 4. Effects of FRS internal modules on the detection performance of different types of ships

Algorithm	Passenger ship	General cargo carrier	Fishing ship	Bulk cargo carrier	Ore carrier	Container ship
FRN	0.654	0.705	0.646	0.696	0.683	0.768
FRN (w/o MFAM)	0.642	0.718	0.633	0.697	0.661	0.752
FRN (w/o AFRM)	0.656	0.701	0.640	0.687	0.684	0.749
FRN (w/o CAM)	0.632	0.693	0.644	0.681	0.678	0.758
FRN (w/o SAM)	0.642	0.703	0.640	0.680	0.680	0.746
FRN (w/o SAM&CAM)	0.637	0.691	0.638	0.680	0.671	0.748
FRN (w/o ARM)	0.641	0.698	0.645	0.696	0.658	0.761
RefineDet	0.622	0.702	0.631	0.678	0.658	0.749

Table 5. Detection results of different algorithms on inshore ship targets

View table

Table 5. Detection results of different algorithms on inshore ship targets

Algorithm	Type	AP	AP^0.5	AP^0.75	AP^s	AP^m	AP^l	FPS
Faster RCNN	two-stage algorithm	0.653	0.919	0.761	0.102	0.423	0.670	6
FPN		0.673	0.941	0.803	0.082	0.472	0.688	16
Libra RCNN		0.683	0.940	0.813	0.072	0.450	0.698	15
YOLOv3	one-stage algorithm	0.509	0.782	0.601	0.122	0.310	0.517	52
RetinaNet		0.659	0.899	0.759	0.201	0.403	0.676	17
FCOS		0.573	0.830	0.673	0.093	0.317	0.585	18
SSD		0.663	0.906	0.793	0.003	0.422	0.687	90
RefineDet		0.674	0.946	0.815	0.134	0.505	0.685	47
Ours		0.692	0.948	0.830	0.122	0.525	0.703	25

Table 6. AP of different algorithms on inshore ship targets

View table

Table 6. AP of different algorithms on inshore ship targets

Algorithm	Type	Passenger ship	General cargo carrier	Fishing ship	Bulk cargo carrier	Ore carrier	Container ship
Faster RCNN	two-stage algorithm	0.591	0.690	0.582	0.644	0.638	0.771
FPN		0.637	0.699	0.614	0.677	0.643	0.769
Libra RCNN		0.638	0.716	0.623	0.686	0.661	0.775
YOLOv3	one-stage algorithm	0.444	0.543	0.443	0.467	0.486	0.674
RetinaNet		0.665	0.685	0.617	0.618	0.594	0.773
FCOS		0.463	0.625	0.533	0.549	0.537	0.731
SSD		0.620	0.712	0.582	0.667	0.628	0.771
RefineDet		0.622	0.702	0.631	0.678	0.658	0.749
Ours(FRN)		0.654	0.705	0.646	0.696	0.683	0.768

Tools

Get Citation

Copy Citation Text

Di Liu, Yan Zhang, Yan Zhao, Zhiguang Shi, Jinghua Zhang, Yu Zhang. Multi-Scale Inshore Ship Detection Based on Feature Re-Focusing Network[J]. Acta Optica Sinica, 2021, 41(22): 2215001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Apr. 28, 2021

Accepted: Jun. 3, 2021

Published Online: Nov. 17, 2021

The Author Email: Zhang Yan (atrthreefire@sina.com)

DOI:10.3788/AOS202141.2215001

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology